1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Persistent Memory Driver 4 * 5 * Copyright (c) 2014-2015, Intel Corporation. 6 * Copyright (c) 2015, Christoph Hellwig <hch@lst.de>. 7 * Copyright (c) 2015, Boaz Harrosh <boaz@plexistor.com>. 8 */ 9 10 #include <linux/blkdev.h> 11 #include <linux/pagemap.h> 12 #include <linux/hdreg.h> 13 #include <linux/init.h> 14 #include <linux/platform_device.h> 15 #include <linux/set_memory.h> 16 #include <linux/module.h> 17 #include <linux/moduleparam.h> 18 #include <linux/badblocks.h> 19 #include <linux/memremap.h> 20 #include <linux/vmalloc.h> 21 #include <linux/blk-mq.h> 22 #include <linux/pfn_t.h> 23 #include <linux/slab.h> 24 #include <linux/uio.h> 25 #include <linux/dax.h> 26 #include <linux/nd.h> 27 #include <linux/mm.h> 28 #include <asm/cacheflush.h> 29 #include "pmem.h" 30 #include "btt.h" 31 #include "pfn.h" 32 #include "nd.h" 33 34 static struct device *to_dev(struct pmem_device *pmem) 35 { 36 /* 37 * nvdimm bus services need a 'dev' parameter, and we record the device 38 * at init in bb.dev. 39 */ 40 return pmem->bb.dev; 41 } 42 43 static struct nd_region *to_region(struct pmem_device *pmem) 44 { 45 return to_nd_region(to_dev(pmem)->parent); 46 } 47 48 static void hwpoison_clear(struct pmem_device *pmem, 49 phys_addr_t phys, unsigned int len) 50 { 51 unsigned long pfn_start, pfn_end, pfn; 52 53 /* only pmem in the linear map supports HWPoison */ 54 if (is_vmalloc_addr(pmem->virt_addr)) 55 return; 56 57 pfn_start = PHYS_PFN(phys); 58 pfn_end = pfn_start + PHYS_PFN(len); 59 for (pfn = pfn_start; pfn < pfn_end; pfn++) { 60 struct page *page = pfn_to_page(pfn); 61 62 /* 63 * Note, no need to hold a get_dev_pagemap() reference 64 * here since we're in the driver I/O path and 65 * outstanding I/O requests pin the dev_pagemap. 66 */ 67 if (test_and_clear_pmem_poison(page)) 68 clear_mce_nospec(pfn); 69 } 70 } 71 72 static blk_status_t pmem_clear_poison(struct pmem_device *pmem, 73 phys_addr_t offset, unsigned int len) 74 { 75 struct device *dev = to_dev(pmem); 76 sector_t sector; 77 long cleared; 78 blk_status_t rc = BLK_STS_OK; 79 80 sector = (offset - pmem->data_offset) / 512; 81 82 cleared = nvdimm_clear_poison(dev, pmem->phys_addr + offset, len); 83 if (cleared < len) 84 rc = BLK_STS_IOERR; 85 if (cleared > 0 && cleared / 512) { 86 hwpoison_clear(pmem, pmem->phys_addr + offset, cleared); 87 cleared /= 512; 88 dev_dbg(dev, "%#llx clear %ld sector%s\n", 89 (unsigned long long) sector, cleared, 90 cleared > 1 ? "s" : ""); 91 badblocks_clear(&pmem->bb, sector, cleared); 92 if (pmem->bb_state) 93 sysfs_notify_dirent(pmem->bb_state); 94 } 95 96 arch_invalidate_pmem(pmem->virt_addr + offset, len); 97 98 return rc; 99 } 100 101 static void write_pmem(void *pmem_addr, struct page *page, 102 unsigned int off, unsigned int len) 103 { 104 unsigned int chunk; 105 void *mem; 106 107 while (len) { 108 mem = kmap_atomic(page); 109 chunk = min_t(unsigned int, len, PAGE_SIZE - off); 110 memcpy_flushcache(pmem_addr, mem + off, chunk); 111 kunmap_atomic(mem); 112 len -= chunk; 113 off = 0; 114 page++; 115 pmem_addr += chunk; 116 } 117 } 118 119 static blk_status_t read_pmem(struct page *page, unsigned int off, 120 void *pmem_addr, unsigned int len) 121 { 122 unsigned int chunk; 123 unsigned long rem; 124 void *mem; 125 126 while (len) { 127 mem = kmap_atomic(page); 128 chunk = min_t(unsigned int, len, PAGE_SIZE - off); 129 rem = copy_mc_to_kernel(mem + off, pmem_addr, chunk); 130 kunmap_atomic(mem); 131 if (rem) 132 return BLK_STS_IOERR; 133 len -= chunk; 134 off = 0; 135 page++; 136 pmem_addr += chunk; 137 } 138 return BLK_STS_OK; 139 } 140 141 static blk_status_t pmem_do_read(struct pmem_device *pmem, 142 struct page *page, unsigned int page_off, 143 sector_t sector, unsigned int len) 144 { 145 blk_status_t rc; 146 phys_addr_t pmem_off = sector * 512 + pmem->data_offset; 147 void *pmem_addr = pmem->virt_addr + pmem_off; 148 149 if (unlikely(is_bad_pmem(&pmem->bb, sector, len))) 150 return BLK_STS_IOERR; 151 152 rc = read_pmem(page, page_off, pmem_addr, len); 153 flush_dcache_page(page); 154 return rc; 155 } 156 157 static blk_status_t pmem_do_write(struct pmem_device *pmem, 158 struct page *page, unsigned int page_off, 159 sector_t sector, unsigned int len) 160 { 161 blk_status_t rc = BLK_STS_OK; 162 bool bad_pmem = false; 163 phys_addr_t pmem_off = sector * 512 + pmem->data_offset; 164 void *pmem_addr = pmem->virt_addr + pmem_off; 165 166 if (unlikely(is_bad_pmem(&pmem->bb, sector, len))) 167 bad_pmem = true; 168 169 /* 170 * Note that we write the data both before and after 171 * clearing poison. The write before clear poison 172 * handles situations where the latest written data is 173 * preserved and the clear poison operation simply marks 174 * the address range as valid without changing the data. 175 * In this case application software can assume that an 176 * interrupted write will either return the new good 177 * data or an error. 178 * 179 * However, if pmem_clear_poison() leaves the data in an 180 * indeterminate state we need to perform the write 181 * after clear poison. 182 */ 183 flush_dcache_page(page); 184 write_pmem(pmem_addr, page, page_off, len); 185 if (unlikely(bad_pmem)) { 186 rc = pmem_clear_poison(pmem, pmem_off, len); 187 write_pmem(pmem_addr, page, page_off, len); 188 } 189 190 return rc; 191 } 192 193 static void pmem_submit_bio(struct bio *bio) 194 { 195 int ret = 0; 196 blk_status_t rc = 0; 197 bool do_acct; 198 unsigned long start; 199 struct bio_vec bvec; 200 struct bvec_iter iter; 201 struct pmem_device *pmem = bio->bi_bdev->bd_disk->private_data; 202 struct nd_region *nd_region = to_region(pmem); 203 204 if (bio->bi_opf & REQ_PREFLUSH) 205 ret = nvdimm_flush(nd_region, bio); 206 207 do_acct = blk_queue_io_stat(bio->bi_bdev->bd_disk->queue); 208 if (do_acct) 209 start = bio_start_io_acct(bio); 210 bio_for_each_segment(bvec, bio, iter) { 211 if (op_is_write(bio_op(bio))) 212 rc = pmem_do_write(pmem, bvec.bv_page, bvec.bv_offset, 213 iter.bi_sector, bvec.bv_len); 214 else 215 rc = pmem_do_read(pmem, bvec.bv_page, bvec.bv_offset, 216 iter.bi_sector, bvec.bv_len); 217 if (rc) { 218 bio->bi_status = rc; 219 break; 220 } 221 } 222 if (do_acct) 223 bio_end_io_acct(bio, start); 224 225 if (bio->bi_opf & REQ_FUA) 226 ret = nvdimm_flush(nd_region, bio); 227 228 if (ret) 229 bio->bi_status = errno_to_blk_status(ret); 230 231 bio_endio(bio); 232 } 233 234 static int pmem_rw_page(struct block_device *bdev, sector_t sector, 235 struct page *page, unsigned int op) 236 { 237 struct pmem_device *pmem = bdev->bd_disk->private_data; 238 blk_status_t rc; 239 240 if (op_is_write(op)) 241 rc = pmem_do_write(pmem, page, 0, sector, thp_size(page)); 242 else 243 rc = pmem_do_read(pmem, page, 0, sector, thp_size(page)); 244 /* 245 * The ->rw_page interface is subtle and tricky. The core 246 * retries on any error, so we can only invoke page_endio() in 247 * the successful completion case. Otherwise, we'll see crashes 248 * caused by double completion. 249 */ 250 if (rc == 0) 251 page_endio(page, op_is_write(op), 0); 252 253 return blk_status_to_errno(rc); 254 } 255 256 /* see "strong" declaration in tools/testing/nvdimm/pmem-dax.c */ 257 __weak long __pmem_direct_access(struct pmem_device *pmem, pgoff_t pgoff, 258 long nr_pages, void **kaddr, pfn_t *pfn) 259 { 260 resource_size_t offset = PFN_PHYS(pgoff) + pmem->data_offset; 261 262 if (unlikely(is_bad_pmem(&pmem->bb, PFN_PHYS(pgoff) / 512, 263 PFN_PHYS(nr_pages)))) 264 return -EIO; 265 266 if (kaddr) 267 *kaddr = pmem->virt_addr + offset; 268 if (pfn) 269 *pfn = phys_to_pfn_t(pmem->phys_addr + offset, pmem->pfn_flags); 270 271 /* 272 * If badblocks are present, limit known good range to the 273 * requested range. 274 */ 275 if (unlikely(pmem->bb.count)) 276 return nr_pages; 277 return PHYS_PFN(pmem->size - pmem->pfn_pad - offset); 278 } 279 280 static const struct block_device_operations pmem_fops = { 281 .owner = THIS_MODULE, 282 .submit_bio = pmem_submit_bio, 283 .rw_page = pmem_rw_page, 284 }; 285 286 static int pmem_dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff, 287 size_t nr_pages) 288 { 289 struct pmem_device *pmem = dax_get_private(dax_dev); 290 291 return blk_status_to_errno(pmem_do_write(pmem, ZERO_PAGE(0), 0, 292 PFN_PHYS(pgoff) >> SECTOR_SHIFT, 293 PAGE_SIZE)); 294 } 295 296 static long pmem_dax_direct_access(struct dax_device *dax_dev, 297 pgoff_t pgoff, long nr_pages, void **kaddr, pfn_t *pfn) 298 { 299 struct pmem_device *pmem = dax_get_private(dax_dev); 300 301 return __pmem_direct_access(pmem, pgoff, nr_pages, kaddr, pfn); 302 } 303 304 /* 305 * Use the 'no check' versions of copy_from_iter_flushcache() and 306 * copy_mc_to_iter() to bypass HARDENED_USERCOPY overhead. Bounds 307 * checking, both file offset and device offset, is handled by 308 * dax_iomap_actor() 309 */ 310 static size_t pmem_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff, 311 void *addr, size_t bytes, struct iov_iter *i) 312 { 313 return _copy_from_iter_flushcache(addr, bytes, i); 314 } 315 316 static size_t pmem_copy_to_iter(struct dax_device *dax_dev, pgoff_t pgoff, 317 void *addr, size_t bytes, struct iov_iter *i) 318 { 319 return _copy_mc_to_iter(addr, bytes, i); 320 } 321 322 static const struct dax_operations pmem_dax_ops = { 323 .direct_access = pmem_dax_direct_access, 324 .dax_supported = generic_fsdax_supported, 325 .copy_from_iter = pmem_copy_from_iter, 326 .copy_to_iter = pmem_copy_to_iter, 327 .zero_page_range = pmem_dax_zero_page_range, 328 }; 329 330 static const struct attribute_group *pmem_attribute_groups[] = { 331 &dax_attribute_group, 332 NULL, 333 }; 334 335 static void pmem_release_disk(void *__pmem) 336 { 337 struct pmem_device *pmem = __pmem; 338 339 kill_dax(pmem->dax_dev); 340 put_dax(pmem->dax_dev); 341 del_gendisk(pmem->disk); 342 343 blk_cleanup_disk(pmem->disk); 344 } 345 346 static int pmem_attach_disk(struct device *dev, 347 struct nd_namespace_common *ndns) 348 { 349 struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev); 350 struct nd_region *nd_region = to_nd_region(dev->parent); 351 int nid = dev_to_node(dev), fua; 352 struct resource *res = &nsio->res; 353 struct range bb_range; 354 struct nd_pfn *nd_pfn = NULL; 355 struct dax_device *dax_dev; 356 struct nd_pfn_sb *pfn_sb; 357 struct pmem_device *pmem; 358 struct request_queue *q; 359 struct gendisk *disk; 360 void *addr; 361 int rc; 362 unsigned long flags = 0UL; 363 364 pmem = devm_kzalloc(dev, sizeof(*pmem), GFP_KERNEL); 365 if (!pmem) 366 return -ENOMEM; 367 368 rc = devm_namespace_enable(dev, ndns, nd_info_block_reserve()); 369 if (rc) 370 return rc; 371 372 /* while nsio_rw_bytes is active, parse a pfn info block if present */ 373 if (is_nd_pfn(dev)) { 374 nd_pfn = to_nd_pfn(dev); 375 rc = nvdimm_setup_pfn(nd_pfn, &pmem->pgmap); 376 if (rc) 377 return rc; 378 } 379 380 /* we're attaching a block device, disable raw namespace access */ 381 devm_namespace_disable(dev, ndns); 382 383 dev_set_drvdata(dev, pmem); 384 pmem->phys_addr = res->start; 385 pmem->size = resource_size(res); 386 fua = nvdimm_has_flush(nd_region); 387 if (!IS_ENABLED(CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE) || fua < 0) { 388 dev_warn(dev, "unable to guarantee persistence of writes\n"); 389 fua = 0; 390 } 391 392 if (!devm_request_mem_region(dev, res->start, resource_size(res), 393 dev_name(&ndns->dev))) { 394 dev_warn(dev, "could not reserve region %pR\n", res); 395 return -EBUSY; 396 } 397 398 disk = blk_alloc_disk(nid); 399 if (!disk) 400 return -ENOMEM; 401 q = disk->queue; 402 403 pmem->disk = disk; 404 pmem->pgmap.owner = pmem; 405 pmem->pfn_flags = PFN_DEV; 406 if (is_nd_pfn(dev)) { 407 pmem->pgmap.type = MEMORY_DEVICE_FS_DAX; 408 addr = devm_memremap_pages(dev, &pmem->pgmap); 409 pfn_sb = nd_pfn->pfn_sb; 410 pmem->data_offset = le64_to_cpu(pfn_sb->dataoff); 411 pmem->pfn_pad = resource_size(res) - 412 range_len(&pmem->pgmap.range); 413 pmem->pfn_flags |= PFN_MAP; 414 bb_range = pmem->pgmap.range; 415 bb_range.start += pmem->data_offset; 416 } else if (pmem_should_map_pages(dev)) { 417 pmem->pgmap.range.start = res->start; 418 pmem->pgmap.range.end = res->end; 419 pmem->pgmap.nr_range = 1; 420 pmem->pgmap.type = MEMORY_DEVICE_FS_DAX; 421 addr = devm_memremap_pages(dev, &pmem->pgmap); 422 pmem->pfn_flags |= PFN_MAP; 423 bb_range = pmem->pgmap.range; 424 } else { 425 addr = devm_memremap(dev, pmem->phys_addr, 426 pmem->size, ARCH_MEMREMAP_PMEM); 427 bb_range.start = res->start; 428 bb_range.end = res->end; 429 } 430 431 if (IS_ERR(addr)) 432 return PTR_ERR(addr); 433 pmem->virt_addr = addr; 434 435 blk_queue_write_cache(q, true, fua); 436 blk_queue_physical_block_size(q, PAGE_SIZE); 437 blk_queue_logical_block_size(q, pmem_sector_size(ndns)); 438 blk_queue_max_hw_sectors(q, UINT_MAX); 439 blk_queue_flag_set(QUEUE_FLAG_NONROT, q); 440 if (pmem->pfn_flags & PFN_MAP) 441 blk_queue_flag_set(QUEUE_FLAG_DAX, q); 442 443 disk->fops = &pmem_fops; 444 disk->private_data = pmem; 445 nvdimm_namespace_disk_name(ndns, disk->disk_name); 446 set_capacity(disk, (pmem->size - pmem->pfn_pad - pmem->data_offset) 447 / 512); 448 if (devm_init_badblocks(dev, &pmem->bb)) 449 return -ENOMEM; 450 nvdimm_badblocks_populate(nd_region, &pmem->bb, &bb_range); 451 disk->bb = &pmem->bb; 452 453 if (is_nvdimm_sync(nd_region)) 454 flags = DAXDEV_F_SYNC; 455 dax_dev = alloc_dax(pmem, disk->disk_name, &pmem_dax_ops, flags); 456 if (IS_ERR(dax_dev)) { 457 return PTR_ERR(dax_dev); 458 } 459 dax_write_cache(dax_dev, nvdimm_has_cache(nd_region)); 460 pmem->dax_dev = dax_dev; 461 462 device_add_disk(dev, disk, pmem_attribute_groups); 463 if (devm_add_action_or_reset(dev, pmem_release_disk, pmem)) 464 return -ENOMEM; 465 466 nvdimm_check_and_set_ro(disk); 467 468 pmem->bb_state = sysfs_get_dirent(disk_to_dev(disk)->kobj.sd, 469 "badblocks"); 470 if (!pmem->bb_state) 471 dev_warn(dev, "'badblocks' notification disabled\n"); 472 473 return 0; 474 } 475 476 static int nd_pmem_probe(struct device *dev) 477 { 478 int ret; 479 struct nd_namespace_common *ndns; 480 481 ndns = nvdimm_namespace_common_probe(dev); 482 if (IS_ERR(ndns)) 483 return PTR_ERR(ndns); 484 485 if (is_nd_btt(dev)) 486 return nvdimm_namespace_attach_btt(ndns); 487 488 if (is_nd_pfn(dev)) 489 return pmem_attach_disk(dev, ndns); 490 491 ret = devm_namespace_enable(dev, ndns, nd_info_block_reserve()); 492 if (ret) 493 return ret; 494 495 ret = nd_btt_probe(dev, ndns); 496 if (ret == 0) 497 return -ENXIO; 498 499 /* 500 * We have two failure conditions here, there is no 501 * info reserver block or we found a valid info reserve block 502 * but failed to initialize the pfn superblock. 503 * 504 * For the first case consider namespace as a raw pmem namespace 505 * and attach a disk. 506 * 507 * For the latter, consider this a success and advance the namespace 508 * seed. 509 */ 510 ret = nd_pfn_probe(dev, ndns); 511 if (ret == 0) 512 return -ENXIO; 513 else if (ret == -EOPNOTSUPP) 514 return ret; 515 516 ret = nd_dax_probe(dev, ndns); 517 if (ret == 0) 518 return -ENXIO; 519 else if (ret == -EOPNOTSUPP) 520 return ret; 521 522 /* probe complete, attach handles namespace enabling */ 523 devm_namespace_disable(dev, ndns); 524 525 return pmem_attach_disk(dev, ndns); 526 } 527 528 static void nd_pmem_remove(struct device *dev) 529 { 530 struct pmem_device *pmem = dev_get_drvdata(dev); 531 532 if (is_nd_btt(dev)) 533 nvdimm_namespace_detach_btt(to_nd_btt(dev)); 534 else { 535 /* 536 * Note, this assumes nd_device_lock() context to not 537 * race nd_pmem_notify() 538 */ 539 sysfs_put(pmem->bb_state); 540 pmem->bb_state = NULL; 541 } 542 nvdimm_flush(to_nd_region(dev->parent), NULL); 543 } 544 545 static void nd_pmem_shutdown(struct device *dev) 546 { 547 nvdimm_flush(to_nd_region(dev->parent), NULL); 548 } 549 550 static void pmem_revalidate_poison(struct device *dev) 551 { 552 struct nd_region *nd_region; 553 resource_size_t offset = 0, end_trunc = 0; 554 struct nd_namespace_common *ndns; 555 struct nd_namespace_io *nsio; 556 struct badblocks *bb; 557 struct range range; 558 struct kernfs_node *bb_state; 559 560 if (is_nd_btt(dev)) { 561 struct nd_btt *nd_btt = to_nd_btt(dev); 562 563 ndns = nd_btt->ndns; 564 nd_region = to_nd_region(ndns->dev.parent); 565 nsio = to_nd_namespace_io(&ndns->dev); 566 bb = &nsio->bb; 567 bb_state = NULL; 568 } else { 569 struct pmem_device *pmem = dev_get_drvdata(dev); 570 571 nd_region = to_region(pmem); 572 bb = &pmem->bb; 573 bb_state = pmem->bb_state; 574 575 if (is_nd_pfn(dev)) { 576 struct nd_pfn *nd_pfn = to_nd_pfn(dev); 577 struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb; 578 579 ndns = nd_pfn->ndns; 580 offset = pmem->data_offset + 581 __le32_to_cpu(pfn_sb->start_pad); 582 end_trunc = __le32_to_cpu(pfn_sb->end_trunc); 583 } else { 584 ndns = to_ndns(dev); 585 } 586 587 nsio = to_nd_namespace_io(&ndns->dev); 588 } 589 590 range.start = nsio->res.start + offset; 591 range.end = nsio->res.end - end_trunc; 592 nvdimm_badblocks_populate(nd_region, bb, &range); 593 if (bb_state) 594 sysfs_notify_dirent(bb_state); 595 } 596 597 static void pmem_revalidate_region(struct device *dev) 598 { 599 struct pmem_device *pmem; 600 601 if (is_nd_btt(dev)) { 602 struct nd_btt *nd_btt = to_nd_btt(dev); 603 struct btt *btt = nd_btt->btt; 604 605 nvdimm_check_and_set_ro(btt->btt_disk); 606 return; 607 } 608 609 pmem = dev_get_drvdata(dev); 610 nvdimm_check_and_set_ro(pmem->disk); 611 } 612 613 static void nd_pmem_notify(struct device *dev, enum nvdimm_event event) 614 { 615 switch (event) { 616 case NVDIMM_REVALIDATE_POISON: 617 pmem_revalidate_poison(dev); 618 break; 619 case NVDIMM_REVALIDATE_REGION: 620 pmem_revalidate_region(dev); 621 break; 622 default: 623 dev_WARN_ONCE(dev, 1, "notify: unknown event: %d\n", event); 624 break; 625 } 626 } 627 628 MODULE_ALIAS("pmem"); 629 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_IO); 630 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_PMEM); 631 static struct nd_device_driver nd_pmem_driver = { 632 .probe = nd_pmem_probe, 633 .remove = nd_pmem_remove, 634 .notify = nd_pmem_notify, 635 .shutdown = nd_pmem_shutdown, 636 .drv = { 637 .name = "nd_pmem", 638 }, 639 .type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM, 640 }; 641 642 module_nd_driver(nd_pmem_driver); 643 644 MODULE_AUTHOR("Ross Zwisler <ross.zwisler@linux.intel.com>"); 645 MODULE_LICENSE("GPL v2"); 646